You are on page 1of 20

This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS 1

Model-Based Diagnosis of Multiple Faults in


Hybrid Dynamical Systems With Dynamically
Updated Parameters
Om Prakash, Arun K. Samantaray, and Ranjan Bhattacharyya

Abstract—In this paper, online fault detection and isola- I. I NTRODUCTION


tion (FDI) of multiple faults for a hybrid dynamical system
AFETY and reliability are key issues in automation of
based on partly decentralized model-based diagnosis architec-
ture is presented. A quantitative model-based diagnosis method
for both discrete and parametric faults occurring sequentially
S modern industries. This calls for very efficient and robust
supervision schemes which perform online fault diagnosis
is proposed here, in which system global model is partitioned and isolation. Various fault diagnosis methods available for
into various subsystems by analyzing the global fault sensitivity this purpose have been classified into quantitative model-
signature matrix (GFSSM) and mode change sensitivity signa-
based, qualitative model-based, and data driven-based meth-
ture matrix (MCSSM) of the global system. A local diagnoser
is designed for each subsystem resulting in local GFSSM and ods. Detailed reviews on these methods are given in [1]–[3].
MCSSM; and supervisory diagnoser is required only when if A recent review on fault diagnosis and fault-tolerant tech-
more than one local diagnoser generate alarm. The proposed niques with model-based and signal-based approaches is
method can also be used to solve the problem of sequentially given in [4] and that for knowledge-based and hybrid/active
occurring faults detection and isolation for a centralized or approaches is presented in [5]. Whatever be the methods,
distributed architecture-based diagnosis. Residual sensitivity sig-
nature with the GFSSM and MCSSM is used to determine a diagnosis approach is mainly classified according to its
a smaller set of possible fault candidates after detection of a fault architecture as centralized, decentralized, and distributed diag-
by the respective local diagnoser. It is shown that use of fault nosis architecture [6], [7]. Most of the developed model-based
direction information from GFSSM and MCSSM improves the fault detection and isolation (FDI) approaches are intended
fault isolation process for discrete or parametric fault and also for continuous systems and consider single fault hypothe-
improves the local parameter estimation process. The identi-
fied discrete fault or estimated parameters are then used to sis. However, large process engineering systems are generally
update the corresponding nominal model which allows for FDI decentralized by nature and are equipped with various dynam-
of the next fault. For robust FDI, adaptive residual threshold is ical components with modern embedded control and software
used which accounts for the process, measurement uncertainties or electronic technologies which may be viewed as combina-
and mode change, including the parameter estimation uncer- tion of continuous and hybrid subsystems. The dynamics of
tainties. Algorithms and codes have been developed in industry
preferred MATLAB-LabVIEW interface for real time monitor-
these systems include both continuous and discrete parts and
ing of the hybrid system. The developed method is first tested on supervision of these systems get complicated with transition
the four-tank hybrid system model through simulation. Then it of modes from one to another. The methods to supervise such
is experimentally validated with real experimental data collected hybrid systems are not firmly established till now in the case
from a reduced-scale equivalent hybrid electronic circuit model of multiple faults.
of a bench mark two-tank system for checking the robustness
Faults in a hybrid dynamical system may be classified into
of the proposed approach. The developed method is shown to
produce accurate diagnosis of sequentially occurring faults in two types [8], [9]: 1) a parametric fault which pertains to
hybrid systems and saves valuable CPU time as compared with degradation of a component and 2) a discrete fault which
centralized approach. pertains to an unexpected mode change. Examples of dis-
Index Terms—Adaptive threshold, bond graph (BG), hybrid crete fault are valve getting stuck in on or off state, relay
system, partly decentralized fault diagnosis, residual, sensitivity fault, pump switching on/off fault, and failure in controller
signature, sequentially occurring faults. command transition, etc. Discrete transitions in a hybrid sys-
tem can be of two types: 1) supervised mode transition (e.g.,
a switch) and 2) autonomous mode transition. The autonomous
mode transition is a discrete event activated by system states.
Manuscript received January 23, 2017; accepted May 24, 2017. This paper On the other hand, supervisor controlled discrete transition
was recommended by Associate Editor C. J. Ferreira Silvestre. (Corresponding refers to switching on or off a device through controller com-
author: Arun K. Samantaray.) mand. Discrete fault refers to the later, i.e., the failure in
The authors are with the Department of Mechanical Engineering,
Systems, Dynamics, and Control Laboratory, Indian Institute of Technology executing a control command to trigger a discrete event either
Kharagpur, Kharagpur 721302, India (e-mail: omimech21@gmail.com; because of a component fault or a command transmission
samantaray@mech.iitkgp.ernet.in; rbmail@ mech.iitkgp.ernet.in). error. The supervised discrete transition is done by software
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. and the control command is known from the software output.
Digital Object Identifier 10.1109/TSMC.2017.2710143 Alternatively, supervisory controller’s model may be used for
2168-2216 c 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

prediction of the controlled mode transitions. In contrast to bounded-output (BIBO) stable and monotone, i.e., for given
supervisor controlled discrete transition, discrete transitions input and output responses, a unique set of parameters exists.
due to autonomous modes are described using the monitored The main contributions of this paper are as follows.
system’s state variables or measurements. The diagnosis mod- 1) A unified diagnosis method for sequentially occurring
ule utilizes the measurement of the plant outputs to compute faults in a hybrid dynamical system is developed. The
the measurable state variables from which the autonomous method utilizes hybrid bond graph (HBG) for model and
mode transitions are predicted. fault diagnoser development. A new idea to update the
So far the research on FDI of hybrid dynamical model after identification of each fault is introduced.
systems [10]–[21] have mostly utilized centralized architec- This approach allows diagnosis of a fault whose symp-
ture. But, large systems are decentralized by nature with toms could have been hidden due to the existence of
various subsystems; for example, power generation, chemi- prior faults.
cal batch process, and water treatment plants. It may not be 2) A partly decentralized diagnoser is proposed and
possible to make the fault diagnoser of a particular subsystem designed by analyzing the global fault sensitivity sig-
fully autonomous and because it may require the information nature matrix (GFSSM) and global mode change sensi-
of some measurements from the other subsystems to maintain tivity signature matrix (GMCSSM) of the system. The
the monitoring and control action of that subsystem. So in this proposed method can also be used to solve the problem
paper, a partly decentralized architecture is used for diagnosis of sequential multiple faults detection and isolation for
of hybrid dynamical systems. Existing centralized, decentral- a centralized or distributed architecture-based diagnosis.
ized, and distributed diagnosis approaches mostly apply to 3) Instantaneous fault sensitivity signatures use the direc-
a single fault case. Hence, the diagnosis framework requires tion of deviation of parametric and discrete faults to
that the isolated fault in a subsystem or system is quickly produce a smaller set of hypothesized faults and thus,
repaired. Sometimes, one fault can trigger a cascade of other permit focused local parameter estimation for improv-
faults and repairing each fault immediately after its occurrence ing the online diagnosis of hybrid system while saving
may be impossible. However, it is possible to tolerate some computational time.
not-so-severe faults and continue the plant operation with the 4) The developed method is not only useful in detecting and
few known faults. However, it should be possible to detect isolating parametric and discrete faults, but also allows
a major or serious fault to take appropriate remedial action mode tracking under pre-existing faults in a system. This
when the process is being operated with pre-existing health method can be used to diagnose abrupt, progressive, and
problems (previously existing and uncorrected minor faults). incipient faults. Intermittent faults can be detected and
FDI methods that use single fault hypothesis fail in such sit- isolated if there is sufficiently long fault duration for
uations because a new fault can conceal or compensate the parameter estimation.
effect of one or more pre-existing faults. A solution can be This paper is organized into seven sections. Section II briefs
implemented by using a lot of sensors so that fault effects the related work in model-based diagnosis. In Section III,
are decoupled from each other. Such a solution is not only hybrid process supervision methodology based on bond
costly but can be impractical because it may not be feasible graph (BG) using partly decentralized architecture is intro-
to measure every process state. Another solution is to build duced. Also, the FDI of sequentially occurring faults using
a bank/set of observers (such as unknown input observers) so the information of residual sensitivities are discussed therein.
that directly immeasurable states or inconsistencies of the plant Section IV discusses the implementation of partly decentral-
can be measured from the observer(s). However, it is difficult ized FDI scheme in MATLAB-Simulink environment and also
to implement observer-based approaches for hybrid and non- shows the validation of developed method for sequential mul-
linear systems because not only building nonlinear observers tiple faults through simulations on a nonlinear hydraulic four-
is difficult, but also maintaining and switching between the tank hybrid system model. Section V discusses the conversion
observers as the hybrid system’s modes change is a very of the hybrid system into a reduced scale equivalent hybrid
challenging task, especially while developing the supervisory electrical/electronic system. Section VI shows results from
system and the associated database. Therefore, use of a fixed experimental validation of developed method through exper-
structure fault signature database and common model for diag- iments performed on the reduced-scale equivalent physical
nosis, as proposed here, is preferred from implementation system. The conclusion is drawn in Section VII.
viewpoint.
This paper addresses the above-mentioned lacunas of sin-
gle fault hypothesis. In the proposed solution, an assumption II. R ELATED W ORK IN FDI AND M OTIVATION
is applied that the chance of faults occurring simultaneously A centralized diagnostic approach is very efficient for small
is too slim and those faults that may appear to be simulta- system because of its simplicity. It uses a global model for
neous are actually separated by a small time interval. The fault diagnosis and control for a system which does not
time interval between two faults is assumed to be of suffi- require any system expansion in future. In this approach, sev-
cient length during which identification of the fault parameter eral sensors installed at various locations directly send the
or mode, as well as, model updating steps are possible. data to the central processing unit which process those data
Furthermore, it is assumed that in a finite time frame, infinite and take the corrective action as a global diagnosis task.
mode transitions cannot occur, the system is bounded-input, However, centralized diagnosis approach has some limitations
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

OM PRAKASH et al.: MODEL-BASED DIAGNOSIS OF MULTIPLE FAULTS IN HYBRID DYNAMICAL SYSTEMS 3

like requirement of large memory size and complicated com- consider sequentially occurring faults under assumption that
munication network, poor scalability, and builds single points any chance of simultaneous faults is very rare.
of failure. So, decentralized or distributed diagnosis is used Among the available FDI methods, quantitative model-
to deal with these problems [6], [7], [22]–[24]. A decentral- based FDI has received much interest recently for diagnosis
ized diagnosis architecture which is scalable to a large system of hybrid dynamical systems. However, this requires an accu-
is composed of local diagnosers whose results are coordi- rate and reliable mathematical model of the real process.
nated by a supervisory diagnoser, i.e., if local diagnosers The unified multienergy domain BG approach [25]–[27] is
are not capable to isolate the actual fault locally then the well-suited for modeling the continuous dynamical systems.
supervisory diagnoser comes into play to refine the diagno- BG tool also is very useful in design and development
sis result [7]. Only in select cases, supervisory diagnoser is of quantitative [28]–[31] and qualitative FDI systems. Both
needed, hence saving valuable CPU time for online monitor- quantitative and qualitative (including artificial intelligence-
ing. This decentralized diagnosis approach may be transformed based) approaches have some commonalities as shown
into distributed diagnosis approach without a supervisory diag- in [32] and [33]. An extension of BG modeling approach to
noser to isolate the actual fault locally [6], [22], [23]; hence, the hybrid dynamical systems termed as HBG [34], [35] is
totally eliminating the supervisory diagnoser at the expense of used in this paper for system modeling. Basic properties of
some additional sensors which are required to totally decou- BG are retained in HBG and thus diagnosis schemes for
ple all subsystems. In [22], BG modeling approach is used hybrid systems have been developed using the common princi-
for local diagnosis of a continuous system. Additional sen- ples for quantitative FDI in [10]–[16] and for qualitative FDI
sors required to achieve local diagnosis can be identified by in [17] and [36].
checking the causal paths of the BG model. In [22], it is Quantitative fault diagnosis method examines the consis-
further shown that supervision of some of the subsystems tency between actual system and its behavior model [37].
with some physical constraints cannot be performed in a fully The inconsistencies of various forms are quantified as resid-
localized way, i.e., decentralized approach is needed. In [6] uals. One of the approaches to real-time quantitative FDI is
and [23], the same BG approach in an event-based frame work through numerical evaluation of analytical redundancy rela-
is used for improving distributed diagnosis for large contin- tions (ARRs) to obtain the residuals. ARRs are expressions
uous systems. In [24], a distributed FDI system is designed for constraints in a system and these are written using system
based on rearrangement of the rows and columns of fault measurements and parameters [28]. When the system behaves
signature matrix (FSM). In [6], it is assumed that either the according to its nominal model, the constraints are satisfied.
different subsystems for a large system are already known Otherwise, the monitored residual trend is analyzed to detect
which need supervision or these may not be known. For and isolate the fault that caused the observed deviation from
a known partitioned system, the main goal of the control the nominal model.
designer is to construct the local diagnosis module that need All residuals should be approximately zero when the sys-
to exchange minimum information among the subsystems. tem is in normal operation, but some of or all the residuals
For unknown partitioned system, the partition structure for if sensitive to a particular fault deviate from zero when
a system and corresponding local diagnosers are created simul- that fault occurs. Residuals are also sensitive to the process
taneously at the design stage in order to ensure that no sensor and measurement uncertainties and these are accounted by
data is exchanged between the subsystems. As suggested using thresholds. A consistent residual is always enveloped
in [6], the following properties must be satisfied in distributed within an estimated threshold which may be fixed [29]–[31]
approach. or adaptive [38]–[40]. The sensitivity of each residual to the
1) All desired single faults can be diagnosed in the system. parametric fault and mode fault are represented by the FSM
2) Each local diagnoser must provide the globally correct and mode change signature matrix (MCSM), respectively. The
faults, i.e., in a faulty condition, only a single local fault signature is represented by a coherence vector C =
diagnoser provides the fault, the others report a null [c1 , c2 , . . . , cn ] whose elements are binary numbers. FSM may
hypothesis. contain structured or unstructured signatures. In structured
3) The exchange and communication of number of sensors FSM, each residual is sensitive to a particular single paramet-
data between the local diagnosers to satisfy the above ric fault only and insensitive to other parametric faults [28].
two conditions is minimal. However, for structured FSM number of fault candidates must
In the proposed approach for sequentially occurring fault diag- be equal to the number of sensors installed in a process which
nosis based on partly decentralized architecture, all properties is not always possible, consequently unstructured or nondi-
as mentioned in [6] are satisfied except the point 2. Instead, agonal type FSM is obtained. In that case, more than one
if more than one local diagnoser generates the fault alarm residual are sensitive to a particular single fault, which means
then only a supervisory diagnoser is required to refine the it may be possible that more than one component share the
actual fault in the system; otherwise, each local diagnoser (as same signature. In [30], a direct numerical method for obtain-
in point 2) is sufficient to isolate the locally occurring faults ing residuals, while bypassing ARR derivation, is proposed
in each subsystem and provide the globally correct faults. The using diagnostic BG (DBG). For diagnosis of hybrid systems,
above literatures consider the single fault hypothesis for con- it has been adapted as diagnostic HBG (DHBG) [11]–[14].
tinuous system. None of the above discussed works consider Since there are transitions of modes in a dynamics of hybrid
multiple faults which is the subject matter of this paper. We system, the ARRs and FSMs need to be derived separately for
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

each mode and detected current mode information needs to be a similar estimation of fault parameter for hybrid systems was
used in FDI algorithm [18], [19], [41]. In DHBG approach, introduced using HBG method in [48].
the constraint relations are derived in global form that can The developed approaches in [20], [45], [47], and [48] are
be valid for every mode; hence, global ARRs (GARRs), suitable for on-line process monitoring because they provide
global FSM (GFSM), and MCSM are introduced [11]–[14]. faster convergence rate for the estimated parameters. However,
The FDI module uses GARRs for residuals evaluation and these methods require estimation of all suspected fault param-
GFSM/MCSM for FDI in a hybrid system. The adaptive resid- eters or PCs, and can be further speed-up if number of fault
ual thresholds are also modified according to the operating candidates to be estimated can be confined to a still smaller set.
mode [41]. In [19], identification of mode in hybrid systems Traditionally in quantitative model-based diagnosis, fault sig-
in the presence of single fault is developed; however, this natures are represented in binary string form whose elements
approach fails to identify the mode in case of multiple faults. are either 0 or 1 depending upon whether the residuals are
In nondiagonal type FSM, parameter estimation technique within threshold or outside the threshold, respectively. These
may be used to isolate any fault influencing unstructured resid- fault signatures are matched to the FSM to detect and isolate
uals. Sensitivity BG (SBG) approach, which computes output faults. The FSM structure decides diagnosability of faults [49].
sensitivity to parameter variations [42], has been proposed to The dimension of fault candidate subspace or PCs can be
speed-up parameter estimation process [43]. However, com- reduced if more information can be embedded in the FSM
plete parameter estimation is time and resource intensive structure. In [21], measured trend of residuals are utilized for
which make it unsuitable for online process monitoring. a hybrid system to generate the fault signature, FSM and
There is always some unavoidable uncertainty in parame- MCSM, which are termed as sensitivity signature, GFSSM and
ter estimation; so rough parameter estimation is sufficient mode change sensitivity signature matrix (MCSSM), respec-
for preliminary fault isolation, but it should be estimated tively. In sensitivity signature, the elements of coherence
quickly before the too much fault propagation [44]. With this vector (C) may contain any of the three terms, i.e., 0, +1, and
aim, parameter estimation based on minimization of error −1 depending upon whether the residual is within threshold,
between model and plant outputs aided by SBG approach crossing an upper threshold and crossing a lower thresh-
and computational load is minimized by identifying possi- old, respectively. The entries in GFSSM and MCSSM are
ble fault candidate subspace and limiting estimation to only modified every time a new measurement is obtained. Since,
the fault candidates [45], [46]. The concept of possible con- GFSSM and MCSSM use residuals’ sensitivities which can
flicts (PCs) introduced in [47] is similar to fault subspace distinguish the signature of increasing or decreasing nature
identification which used an integrated qualitative–quantitative of the value of parameters and mode, the richer informa-
approach based on HBG for diagnosis of single abrupt fault. tion structure gives better fault isolation ability compared
In [47], a minimal estimator is found by decomposing the sys- to standard FSM and MCSM for hybrid system. However,
tem model into smaller independent subsystems for the local even this method is incapable of isolating the multiple faults
parameter estimation task. This approach starts with qualita- because pre-existing fault symptoms may be compensated or
tive fault isolation based on temporal causal graph (TCG) to concealed by those of the newly appearing faults. Recently
minimize possible fault hypotheses, and then uses a focused in [50], a DHBG approach is used for sequential multiple fault
quantitative parameter estimation scheme to identify the true diagnosis for hybrid system and composite harmony search
fault. The detection of fault is based on comparing the out- technique is used for the fault identification. The fault iden-
put of a hybrid observer with the measured sensor’s value tification approach proposed therein introduces a second-level
by incorporating a Kalman filter for continuous tracking and residual (SLR) when already a fault is present. Unfortunately,
a mode change detector. It is assumed therein that discrete SLR is derived from time derivatives of GARRs. While
modes are known and the focus was on parametric fault maximum of two faults are considered in [50], the SLRs
only. The fault signature was generated qualitatively by using become noisier because of time derivative as more faults
TCG model and was improved using the deviation of residual appear. Moreover, the method in [50] is applied to open-loop
magnitude and slope. When a fault occurred in the system, systems only.
the hybrid observer was unable to track autonomous mode Since, most of the developed methods based on central-
transition as model is invalid after fault. Consequently, fault ized architecture have some limitations as discussed earlier.
hypotheses are made by using past history of mode informa- To eliminate those limitations, decentralized or distributed
tion up to the level of system diagnosability. Thus, autonomous methods were developed which are intended for large contin-
modes are not identified based on their continuous dynam- uous systems only. Therefore, an improved diagnosis method
ics, but only hypothesized from the mode change conditions. is proposed here which can deal with sequentially occurring
Furthermore, it was assumed that during data collection for faults based on partly decentralized architecture for hybrid
fault parameter estimation, there is no autonomous mode systems. This method takes advantage of sensitivity signa-
change. Discrete faults were not at all considered in [47]. This ture matrix structure to reduce number of fault candidates or
parameter estimation approach was further extended to hybrid PCs and uses online estimation of fault parameters followed
systems in [20] for multiple fault diagnosis. However, the by model updating to continuously diagnose sequentially
assumptions applied in [20] and [45] to reduce parameter esti- occurring faults.
mation resource overload have a restriction that two or more Robustness in FDI is a very important issue because process
fault effects should not be canceling each other. Likewise, and measurement uncertainties may give some false alarms
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

OM PRAKASH et al.: MODEL-BASED DIAGNOSIS OF MULTIPLE FAULTS IN HYBRID DYNAMICAL SYSTEMS 5

and output variables (Y) of the plant. GARR of a hybrid


subsystem or system with uncertainty can be expressed as [14]

GARRn ± λ = 0 (1)

where MD = [a1 , a2 , . . . , ak , . . . , am ]T denotes the con-


trolled junction mode vector having m controlled junction
mode variables, θ = [θ1 , θ2 , . . . , θj , . . . , θp ]T is a param-
Fig. 1. Partly decentralized architecture for fault diagnosis.
eter vector of known nominal values having p number of
parameters, U denotes known inputs, and Y denotes mea-
sured variables. Also, GARRn denotes a nominal part and λ
denotes an uncertain part of GARR. The usual approach to
and/or misdetections. Robust FDI methods minimize false
identify inconsistencies due to fault is to check whether the
alarms and misdetections by considering process and mea-
numerically evaluated result from GARRn remains bounded
surement uncertainties. There are two methods to achieve
between ±λ at any mode. The residual (r) and adaptive thresh-
robustness in FDI. The first is active approach and second
old (ε = |λ|) are obtained through numerical evaluation of
is passive approach [28]. Choice of the approach depends
GARRn and uncertain part λ, respectively, by using the sys-
upon the specific system. In this paper, a passive approach
tem’s measurements (U, Y), mode information (MD), and its
has been utilized [10], [28], [38]–[40] with an adaptive resid-
nominal parameters (θ ). The LFT form of DBG (or DHBG)
ual threshold scheme. Generation of adaptive threshold needs
can be used to numerically evaluate the residuals (r) and
specification of both parameter and measurement uncertain-
corresponding adaptive thresholds (ε) [10], [28], [39], [40].
ties. For hybrid systems, the adaptive threshold varies with
A coherence vector (C) represents the signature for a fault.
the transition of mode [41]. Furthermore, with continuous
For n residuals, C = [c1 (t), c2 (t), . . . , cn (t)], where ci (t) =
parameter estimation with variable degree of estimation error,
(ri (t)), (i = 1, 2, . . . , n) are the alarm states and (ri (t))
the adaptive threshold generation algorithm needs continuous
is a decision procedure which tests each residual ri (t) against
updating.
a time-varying (adaptive) residual threshold εi (t)
In summary, this paper presents an improved and easy to

implement FDI scheme for sequentially occurring faults by ⎨ 0, if − εi (t) ≤ ri (t) ≤ εi (t)
using partly decentralized architecture for hybrid dynamic ci (t) = (ri (t)) = +1, if ri (t) ≥ εi (t) (2)

systems based on GARRs evaluation, GFSSM/MCSSM −1, otherwise.
containing fault sensitivities, continuous estimation of fault
parameters with SBGs and adaptive threshold evolution with During online supervision, alarm states are obtained at
linear fractional transformation (LFT), and model updating each sampling instance and the fault alarm is issued when
and combination of some new ideas. This paper includes the coherence vector C = [0, 0, . . . , 0]. Upon an alarm or
experimental validation of the proposed approach. fault detection, C is matched with rows of a FSM and any
unique match with fault signature of a component/parameter
can isolate the actual fault [14], [28]. A component fault is
III. M ODEL -BASED D IAGNOSIS M ETHODOLOGY monitorable or detectable (represented by monitorability index
A. Basic Framework for Partly Decentralized Mb = 1) if one or more residuals is/are sensitive to it. A fault
Fault Diagnosis can be isolated (represented by isolability index Ib = 1) if Mb
= 1 for it and its fault signature uniquely matches with a row
The basic architecture of partly decentralized fault diagno-
of FSM.
sis as proposed in this paper is composed of various local
smart diagnosers as shown in Fig. 1. If more than one local
diagnoser generates the fault alarm then only the supervisory B. FSM, GFSM, GFSSM and MCSM, and MCSSM
diagnoser comes into play to refine the actual fault in a sys- A FSM, S, represents the qualitative influence of the set of
tem. Each local diagnoser is sufficient to isolate the locally faults and on the corresponding fault signatures. It is matrix
occurring faults in each subsystem and provide the globally whose elements are either 1 or 0 as determined from
correct faults. 
1, if ri is a function of θj
This paper considers a class of dynamical systems which Sji = (3)
0, otherwise
contains family of continuous dynamical system in which
each continuous dynamics is triggered by a discrete mode where ri is the residual corresponding to the ith column of
(supervisory controller mode or autonomous mode) accord- FSM, θ j is the component or parameter corresponding to the
ing to certain conditions. For example: embedded system, jth row of FSM, i ∈ {1, 2 . . . n}, j ∈ {1, 2 . . . p}, and n and p
switched electrical circuit, hydraulic system with on–off con- are, respectively, the number of residuals and parameters.
troller used for pumps and valves, etc. For more examples of Dynamics of a hybrid system includes continuous-time and
hybrid dynamical systems, readers may refer [51]. discrete-time dynamics. Therefore, the FSM for a hybrid
For hybrid systems, GARRs are used in places of ARRs. system depends on the discrete mode and generally, a set
GARR(MD, θ , U, Y) is a constraint expressed in terms of of FSMs are constructed corresponding to various combina-
mode information (MD), nominal parameters (θ ), inputs (U), tions of on/off status of discrete modes. Instead of several
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

FSMs, a GFSM, GS, can be used to form a combined


representation [11]–[14]. Elements of GS are determined from


⎪ f (a1 . . . am ), if ri is a function of θj depends

on the values a1 . . . am
GSji =

⎪ 1, if ri is a function of θj at all modes
⎩ Fig. 2. Capacitance (C) element in derivative causality modeled in LFT
0, otherwise
form.
(4)
where f is a logical function of the switched junction and/or
source, ak ∈ [0, 1], k = 1 . . . m indicates the discrete mode of
a switched/controlled junction, and m is the total number of
discrete modes.
GFSM can be extended to GFSSM so that the influence of
increasing (θj ↑) and decreasing (θj ↓) trends of a faulty param- Fig. 3. LFT form model of resistive (R) element in conductive causality.
eter value on the residuals can be differentiated. Elements of
GFSSM change with time (as per system mode) and they The residual threshold is selected in such way that the resid-
depend on the sign of residual’s sensitivity to change in the ual response remains bounded within it for known ranges
parameter values at a given instance [21]. The elements (GSS) of process and measurement uncertainties. Adaptive thresh-
in the GFSSM are obtained as olds using LFT method [10], [28], [39], [40] are used here, in
⎧   which system uncertainties in parameter values are separated

⎪ −sign ∂ri /∂θj , if ri is a function of θj

⎪ from the nominal parameter-based model and the possible

⎪ and θj is expected to

⎪ deviations in parameter values are modeled as feedback loops
⎨   increase due to fault
of internal variables. For example, when the real parameter
GSSji = +sign ∂ri /∂θj , if ri is a function of θj (5)

⎪ value of a capacitance C is not accurately identified, it can be

⎪ and θj is expected to

⎪ expressed as Cn ± C = Cn (1 ± δC ), where Cn is represented

⎪ decrease due to fault
⎩ as nominal parameter value and ±C = ±δC Cn is the mul-
0, otherwise.
tiplicative uncertainty part of the parameter. If the C element
Faults in discrete modes also affect residuals and this influ- is modeled in derivative causality then its constitutive relation
ence is qualitatively captured by MCSM whose elements are is as
represented as 1 1   ė
 f = ė = 1 ∓ δ1/C ė = ∓ w1/C (8)
MCSMki =
1, if ri is a function of ak
(6) Cn ± C Cn Cn
0, otherwise.
where ∓(δ1/C /Cn )ė = ∓w1/C is additional flow injection
MCSM can be extended to form MCSSM so that increas- which accounts for the bounded uncertainty in the parame-
ing (ak ↑) and decreasing (ak ↓) discrete mode faults (transition ter value. This additional flow injection can be thought of as
from 0 to 1 or vice versa) can be differentiated. The ele- a unknown disturbance. Here, δ1/C is the error bound or uncer-
ments in MCSSM are time-varying and they are computed tainty while the parameter value 1/C is estimated. Fig. 2 shows
at each sampling step by utilizing the qualitative sensitivity of LFT BG form modeling of this way of representation of
residuals to mode changes [21] given as uncertainty in capacitance (C-element) [10], [28], [39], [40].
⎧ Likewise, the constitutive equation for an uncertain nonlin-

⎪ −sign(∂ri /∂ak ), if ri is a function of ak

⎪ ear R element representing fluid flow through a pipe or valve

⎪ and ak is expected to

⎪ can be expressed (for conductive causality in R) as
⎨ change from 0 to 1
MCSSMki = +sign(∂ri /∂ak ), if ri is a function of ak 1 √ 1 √

⎪ f = . e= 1 ∓ δ1/R e

⎪ and ak is expected to Rn ± R Rn

⎪ √  √ √

⎪ change from 1 to 0 = (Cd ∓ Cd ) e = Cd 1 ∓ δ1/R e = Cd e ∓ w1/R
⎩ (9)
0, otherwise.
(7) where δ1/R = Cd /Cd , e is the pressure drop across the
valve/ pipe, Cd is the nominal valve discharge coefficient value
This form of fault signature is called sensitivity signature. and Cd is the error bound
For simple systems, these signatures can be determined ana- √ or uncertainty in estimating Cd ,
and (∓w1/R ) = ∓δ1/R Cd e is an additional flow injection to
lytically and expressed as closed form relations. If GARRs account for the parameter estimation error. Fig. 3 shows LFT
cannot be expressed in analytical closed form then numerical BG form modeling of the constitutive relations of uncertain R
sensitivities can be computed using SBG [43], [45], [52]. element given in (9) [10], [28], [39], [40]. Similarly, uncer-
tainties in other parameter values (I-element, and TF and GY
C. Adaptive Thresholds for Robust Diagnosis two-ports) can be modeled in LFT BG form.
Adaptive time-varying residual thresholds help to reduce Also, the measurement uncertainties (sensor noises) can be
misdetections of faults and false alarms. This is the simplest decoupled from its nominal part by using the BG modeling
method for achieving robustness in the process supervision. technique as proposed in [53] and [54]. The sensors used in
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

OM PRAKASH et al.: MODEL-BASED DIAGNOSIS OF MULTIPLE FAULTS IN HYBRID DYNAMICAL SYSTEMS 7

At each local smart station, all evaluated residuals should


be approximately zero, i.e., lie within the adaptive thresh-
olds, during normal operation of system. When the value of
a parameter (say θj ) of θ deviates sufficiently (outside its
uncertainty bound) or a mode of MD (say ak ) deviates from
its nominal behavior then the residuals sensitive to the devi-
ation of θ j or ak stray outside the residual thresholds at the
corresponding local station.
When one or more residuals violate their thresholds, it is
initially hypothesized that the alarm is consequence of a dis-
crete mode fault. Then the sensitivity signature at that instance
(+1, −1, and 0, respectively, for crossing upper and lower
thresholds and for remaining within thresholds) are determined
and matched with MCSSM at the respective local station.
If a unique match is found then the discrete mode ak is
identified as the source of fault. If the sensitivity signature
does not have unique match in MCSSM then the ARR-based
mode tracking [14] procedure is applied. In this procedure,
all GARRs corresponding to the inconsistent residuals are
re-evaluated for each mode fault hypothesis and any such
Fig. 4. Flow chart of sequential multiple faults diagnosis for a hybrid system. combination that gives consistent residuals is used to iden-
tify the faulty modes. Once a mode fault has been identified,
BG framework are flow detector (Df) and effort detector (De).
it is assumed to persist thereafter. The identified discrete mode
In DHBG framework, these are dualized as modulated sources
fault ak is then used to update the DHBG model, GARRs and
of flow (MSf) and effort (MSe), respectively. The measure-
the thresholds so that subsequent faults can be monitored. The
ment uncertainties, treated as additive uncertainties, are then
initial mode of the system, if unknown, can be determined by
represented as
using the ARR-based mode tracking procedure [14].
MSf = MSfn ± MSf If the initial hypothesis of discrete mode fault is found to
MSe = MSen ± MSe be invalid then the new hypothesis of only single paramet-
ric fault is considered. If the sensitivity signature determined
where MSfn and MSen are the nominal part and MSf and from residual threshold violations uniquely matches with a row
MSe are the error in the respective measurements. (corresponding to parameter θ j ) of GFSSM at the respec-
In this paper, it is assumed that uncertainties in the various tive local station then the faulty component is isolated. If
parameters are uncorrelated. Therefore, the adaptive thresholds two or more components have equal signature in the GFSSM
are implemented by assuming possible maximum deviations in at respective local station then targeted parameter estimation
each of the parameters. As a consequence, the adaptive thresh- techniques, which try to promptly estimate only a small set
olds are a little relaxed and mostly avoid false alarms. The of possible fault parameters or PCs, can be used to isolate the
implementation of measurement uncertainties requires addi- actual fault candidate [20], [45], [47], [48]. Parameter estima-
tional static threshold over the adaptive threshold. Since the tion also gives the severity of the fault and can be used
adaptive threshold is already relaxed, we do not include the in prognosis to determine the remaining useful life of the
further static threshold because that can lead to further detec- faulty component. Now the component containing detected
tion delay. On the other hand, some spikes that may appear in fault θ j is finally declared as a faulty component. The esti-
residuals due to noisy measurements are removed through low- f
mated parameter value represented is by θj . If the nominal
pass filtering of the residuals (with small filter time constant)
part of each GARR is evaluated again with the estimated
both in the simulations and experiments.
parameter and updated adaptive thresholds then the corre-
D. Multiple Fault Diagnosis Method sponding residual does not cross the any threshold [20]. So,
The flow chart for sequential multiple faults diagno- the original θ vector is updated by replacing the nominal θ j
f
sis of hybrid dynamical system proposed in this paper is by the estimated parameter value θj of the real subsystem.
f
shown in Fig. 4. In a hybrid system, the general form of Now updated θ = [θ1 , θ2 , . . . ,θj , . . . , θP ]T , is assumed to
GARR at a local smart station obtained from a DHBG be the new nominal parameter vector that is used to update
model is given as GARR = GARR(MD, θ , U, Y), where DHBG model in LFT form. Evaluation of nominal part of
MD = [a1 , a2 , . . . , ak , . . . , am ]T denotes the controlled junc- each updated GARR using the faulty plant measurements,
tion mode vector having m controlled junction mode variables, known inputs, updated parameter vector θ , and the known
θ = [θ 1 , θ2 , . . . , θj , . . . , θp ]T denotes a known parameter vec- mode information shows residuals values lying within updated
tor having p number of the corresponding smart station’s adaptive thresholds. If any one of the nominal parameter of
nominal parameters, U denotes known inputs, and Y denotes updated θ (more than uncertainty value) changes or any new
measured variables. Online evaluation of nominal part of each inconsistency in updated MD occurs, again only a set of partic-
GARR using MD, θ , U, and Y provides residual. ular residuals, which are sensitive to this inconsistency, cross
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

the adaptive threshold. In the similar manner as discussed,


a fault is detected and isolated and the next component related
to abnormal parameter or discrete mode fault is declared as
a faulty.
In the existing diagnosis methods without updating the fault
parameter or fault mode and threshold, later fault effects
may be concealed or compensated by the effects of other
pre-existing faults. In [55], a high-gain observer-based adap-
Fig. 5. Schematic of a four-tank hybrid system.
tive identification method is proposed to simultaneously detect
the abrupt deviation in the parameter magnitudes at differ-
ent time points and their estimates over the resulting time
intervals. Instead of constant threshold, multiple adaptive
thresholds to detect and identify a series of abrupt changes in
system parameter values are proposed. In effect, the thresholds
are widened as the fault magnitude increases. In this paper,
however, the model is updated to reduce residual magnitudes
and the thresholds remain the same.
Thus, this method is suitable for online monitoring of hybrid
systems in case of sequentially occurring faults wherein less
serious faults can be tolerated and maintenance activities can
be scheduled selectively for serious faults. Note that the devel-
oped method, though demonstrated in the following sections
through simulations and experiments conducted with abrupt
parametric faults, is equally applicable to incipient and pro-
gressive faults. The method can detect any intermittent fault
and can isolate it if the fault magnitude and duration is
sufficient for parameter estimation.
Note that after isolation of actual fault (discrete mode fault
or parametric fault) and its severity, a fault tolerant con-
Fig. 6. BG model of the two tank hybrid system.
trol (FTC) system takes the decision whether the fault can
be accommodated or system reconfiguration is required [28].
In [56], an observer-based control algorithm is presented for
sensor fault diagnosis in systems with multiple state and input also used. The whole system is broken into two local smart
time delays, so that estimated state and sensor data can be used stations S1 and S2 .
for FTC of the system. Designing a FTC system is one of the This system exhibits hybrid dynamics of various kinds.
other challenging areas for hybrid dynamical system. Due to When water level in tank T 1 exceeds level H L1 then the
large parametric fault or discrete mode fault in a subsystem, water starts flowing from tank T1 to tank T2 through drainage
the whole system may become unstable or stable depending pipe L1 . Similarly, when water level in tank T2 exceeds level
upon the severity of the fault [57], [58]. We consider a sys- H L2 then the water starts flowing from tank T2 to atmo-
tem that remains BIBO stable and output controllable under sphere through the drainage pipe L2 . These hybrid dynamics
all possible parametric and mode faults. termed here as autonomous mode transitions as these tran-
sitions depend upon state variables (liquid levels in tanks
T1 and T2 ). Also the valves V1 , V3 , and V4 considered here
IV. A PPLICATION TO D ISTRIBUTED have one of two discrete states, i.e., either on or off. The dis-
F OUR -TANK S YSTEM crete state of valve V4 (aV4 ) is governed by time of filling tank
The four-tank hybrid system is adapted from two-tank T4 ; after every 500 s, the tub is replaced by another tub of sim-
hybrid system [10] for testing and also to show the scalability ilar size and the time taken to replace a tub is 300 s and valve
of our proposed method. The process and instrumentation dia- V4 is in closed condition only during the replacement time of
gram and BG model are shown in Figs. 5 and 6, respectively. tub (state of aV4 is shown in Fig. 7). To introduce leakage
The considered system contains four tanks T1 , T2 , T3 , and T4 fault in corresponding tanks T1 , T2 , T3 , and T4 , four imagi-
which are connected with various pipes and valves V1 , V2 , V3 , nary valves having coefficients of discharge CdLeak1 , CdLeak2 ,
and V4 , whose coefficients of discharge are CdV 1 , CdV 2 , CdV 3 , CdLeak3 , and CdLeak4 , respectively, are modeled. For measur-
and CdV 4 , respectively. This system also has a hydraulic pump ing the water levels in tanks T1 , T2 , T3 , and T4 and water
controlled by a PI-controller which regulates the liquid level flow input by pump, four level sensors H 1 , H 2 , H 3 , and H 4
in tank T 1 . The input flow (QP ) to the tank T1 from the pump and one flow sensor QP are installed in the system. The atmo-
is proportional to the output of the PI-controller (U PI ). Two spheric pressure is assumed to be reference pressure and taken
drainage pipes (L1 and L2 ), which show linear behavior with as zero. The inertia effect of water flow, small angle γ and
coefficients of discharge are CdL1 and CdL2 , respectively, are volume of liquid in the pipe are neglected in this system; with
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

OM PRAKASH et al.: MODEL-BASED DIAGNOSIS OF MULTIPLE FAULTS IN HYBRID DYNAMICAL SYSTEMS 9

Fig. 7. Discrete state of mode aV4 which repeats with a time period of
800 s.

the latter two may be treated as part of the uncertainties in esti-


mating the capacities of tanks. Such assumption where minor
system dynamics is neglected reduces the computational load
and thereby, increases the fault diagnosis speed and simplifies
the supervision system implementation.
Pump saturation characteristic (
P ) and PI-controller output
law (
PI ) are given as in (10) and (11), respectively. These
constitutive laws have been used in [28], [29]

⎨ UPI , if 0 ≤ UPI ≤ fmax
QP = 0, if UPI ≤ 0 =
P (UPI ) (10)

fmax , if UPI > fmax

   
UPI = KP Spt − ρ.g.H1 (t) + KI Spt − ρ.g.H1 (t) dt
=
PI (H1 (t)). (11)
In (10) and (11), f max is the maximum flow from the pump,
Spt is a pressure (or level) set point, and KP and KI are the
Fig. 8. DHBG model of the four tank hybrid system in LFT form.
proportional and integral gains, respectively.
The BG in Fig. 6 follows usual conventions [28]. The vir-
tual sensor Df’: f c4 paired with modulated flow source Msf: √

f c4 model the drainage pipe between tanks T1 and T2 . Note GARR4 : aV1 Cdv1 ρg H1 (t) − H2 (t)sign(H1 (t) − H2 (t))
that one-junction with subscripts C1 , C2 , . . . , C5 are controlled d
− CT2 (ρgH2 (t)) + a1 CdL1 ρg · (H1 (t) − HL1 )
junctions which are associated with discrete modes. dt
The ARRs for the actuators and the controllers are simply − a2 CdL2 ρg(H2 (t) − HL2 )

obtained from comparisons of the respective input and output − CdV2 |ρg(H2 (t) − H3 (t))|sign(H2 (t) − H3 (t))
relationships as

− CdLeak2 |ρgH2 (t)| ± λ4 = 0 (15)


ARR1 : QP −
P (UPI ) = 0 (12) √

GARR5 : CdV2 ρg |H2 (t) − H3 (t)|sign(H2 (t) − H3 (t))


ARR2 : UPI −
PI (H1 (t)) = 0. (13)

d
− CT3 (ρgH3 (t)) − aV3 CdV3 |ρg(H3 (t) − H4 (t))|
The above mentioned ARRs lead to diagonal entries in the dt

FSM (structured residuals [37]) and hence the associated faults × sign(H3 (t) − H4 (t)) − CdLeak3 |ρgH3 (t)| ± λ5 = 0
are uniquely isolated. In the event of such faults, the plant
(16)
needs to be shut down; so we do not consider such faults

in our further analysis. Sensor faults are readily isolated with GARR6 : aV3 CdV3 |ρg(H3 (t) − H4 (t))|
additional hardware redundancy. ARRs for redundant sensors d
× sign(H3 (t) − H4 (t)) − CT4 (ρgH4 (t)) − aV4 CdV4
are algebraic, statistical, or empirical relations between mea- dt

surements from various sensors which check consistency of × |ρgH4 (t)| − CdLeak4 · |ρgH4 (t)| ± λ6 = 0 (17)
the measurements. Several works have been reported for sen-
sor failure identification and they can be readily applied in where CTi = Ai /g, (i = 1, 2, 3 and 4), g is gravity
the hybrid system considered here [28], [29]. Therefore, we  
do not emphasize on sensor fault detection in this paper. 0, if H1 (t) ≤ HL1 0, if H2 (t) ≤ HL2
a1 = , a2 =
The DHBG model of the four-tank hybrid system in LFT 1, if H1 (t) > HL1 1, if H2 (t) > HL2
form is shown in Fig. 8. Four virtual flow sensors (Df*) are
used to derive four constraints GARR3 , GARR4 , GARR5 , and ρ is fluid density, A is tank cross section area with its sub-
GARR6 as follows: script indicating the respective tank, and aV1 , aV3 , and aV4
d √ indicate on or off mode of valves V1 , V3 , and V4 , respec-
GARR3 : QP − CT1 (ρgH1 (t)) − aV1 CdV1 ρg tively. Also, (14)–(17) contain uncertain parts λi of GARRi

dt
(i = 3, 4, 5, and 6). The influences of uncertainties in
× |H1 (t) − H2 (t)|sign(H1 (t) − H2 (t)) − a1 CdL1 ρg

different parameters on a GARR are uncorrelated with pos-
× (H1 (t) − HL1 ) − CdLeak1 |ρgH1 (t)| ± λ3 = 0 sibility of canceling out each other. Therefore, absolute values
(14) of the individual influences are taken for adaptive threshold
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

evaluation as follows:

d
λ3 = δCT1 · CT1 · (ρ · g · H1 (t))

dt



+ aV1 · δCdV1 · CdV1 · ρ · g · |(H1 (t) − H2 (t))|

+ a1 · δCdL1 · CdL1 · ρ · g · (H1 (t) − HL1 )



+ δCdLeak1 · CdLeak1 · |ρ · g · H1 (t)| (18)



λ4 = aV1 · δCdV1 · CdV1 · |ρ · g · (H1 (t) − H2 (t))|

d
+ δCT2 · CT2 · (ρ · g · H2 (t))
dt

+ a1 · δCdL1 CdL1 · ρ · g · (H1 (t) − HL1 )

+ a2 · δCdL2 · CdL2 · ρ · g · (H2 (t) − HL2 )



+ δCdV2 · CdV2 · |ρ · g · (H2 (t) − H3 (t))|


Fig. 9. Architecture for fault diagnosis of four-tank system.
+ δCdLeak2 · CdLeak2 · ρ · gH2 (t) (19)


TABLE I
λ5 = δCdV2 · CdV2 · |ρ · g · (H2 (t) − H3 (t))| GFSM (GS) FOR THE H YBRID F OUR TANK S YSTEM

d
+ δCT3 · CT3 · (ρ · g · H3 (t))
dt



+ aV3 · CdV3 · |ρ · g · (H3 (t) − H4 (t))|



+ δCdLeak3 · CdLeak3 · |ρ · g · H3 (t)| (20)



λ6 = aV3 · δCdV3 · CdV3 · |ρ · g · (H3 (t) − H4 (t))|

d
+ δCT4 · CT4 · (ρ · g · H4 (t))
dt



+ aV4 · δCdV4 · CdV4 · |ρ · g · H4 (t)|



+ δCdLeak4 · CdLeak4 · ρ · gH4 (t) . (21)
TABLE II
Using (4) and (6) on (14)–(17), the GFSM and MCSM MCSM FOR THE H YBRID F OUR TANK S YSTEM
for four-tank hybrid system are presented in Tables I and II,
whereas using (5) and (7) on (14)–(17), the GFSSM and
MCSSM are given in Tables III and IV.
The parameters related to leakage fault in tank Ti , i.e.,
CdLeaki , (i = 1, 2, 3, and 4), have only increasing possibility,
while for other [i.e., CdV i , (i = 1, 2, 3, 4), CdL1 , CdL2 ], both
increasing (i.e., leakage) and decreasing faults (i.e., blockage)
are possible and discrete stuck on and off fault for aV1 , aV3 , based on arranging the rows and columns of FSM can be found
and aV4 are possible. These possibilities are considered in in [24]. Under limited sensors, all single faults of interest
Tables III and IV. In these tables, a fault is structurally iso- are not possible to isolate. In this paper, two local smart
latable if Ib = 1 which requires that Mb = 1. If a fault is work stations S1 and S2 are designed to diagnose the faults
not isolatable (Ib = 0) due to two or more fault signatures in subsystems using GFSM/MCSM and GFSSM/MCSSM
becoming same then the fault parameter is estimated if it is (Tables I–IV). Only single local diagnoser provides the fault
observable, i.e., Mb = 1. Thus, the global condition for fault for a faulty subsystem, the others report a null hypothesis [6].
diagnosability is defined as Mb = 1. When two or more local diagnosers provide the faults, it means
The scaled-up four-tank hybrid system adapted from two- supervisor is required to solve the conflict and isolate the
tank hybrid system [10] is partitioned by reviewing the actual faults. For instance, refer to Tables I–IV in which zone
GFSM/MCSM and GFSSM/MCSSM (Tables I–IV) of global of subsystem S1 is shown within dotted lines and of subsys-
system as S1 and S2 as shown in Fig. 9. The partitioned sub- tem S2 is shown within dashed lines. The signature of CdV2
systems are selected in such a way that each local diagnoser is shared by both subsystems and when the fault in valve V2
must provide the globally correct faults in a faulty condition occurs then both smart stations S1 and S2 generate alarms
with minimum exchange of information to supervisory diag- and supervisory diagnoser is needed. In all other cases, local
noser. Similar approach for designing the distributed diagnosis diagnosis is performed which saves valuable CPU time for
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

OM PRAKASH et al.: MODEL-BASED DIAGNOSIS OF MULTIPLE FAULTS IN HYBRID DYNAMICAL SYSTEMS 11

TABLE III
GFSSM (GSS) FOR THE H YBRID F OUR -TANK S YSTEM

Fig. 10. Simulink model of four-tank hybrid system with provisions to


introduce faults; where K1 = Cdv1 , K2 = Cdv2 , K3 = CdL1 , K4 = CdL2 ,
K5 = CdLeak1 , K6 = CdLeak2 , K7 = 1/ρg, K8 = g/A1 , K9 = g/A2 ,
K10 = Cdv3 , K11 = Cdv3 , K12 = CdLeak3 , K13 = CdLeak4 , K14 = g/A3 , and
K15 = g/A4 .
TABLE IV
MCSSM FOR THE H YBRID F OUR -TANK S YSTEM

For the validation of the proposed method, the HBG model


is converted into MATLAB-Simulink model (with provisions
of faults introduction) to generate the process data through
simulation as shown in Fig. 10. To introduce faults in the
valves V1 (Cdv1 ) and V3 (Cdv3 ), leakage in T1 (CdLeak1 ), and
discrete fault in av4 , step functions multiplying blocks as high-
lighted in Fig. 10 are used. Switch blocks are used to model
autonomous transitions. Similarly, the DHBG model in LFT
form may be converted into MATLAB-Simulink model for
residuals and adaptive thresholds evaluation [10].
Here, supervision of subsystem S1 is done independently
in a separate computer. The residuals and thresholds eval-
online monitoring. If possible, local parameter estimation will uation are done in a MATLAB program that evaluates the
be used to isolate the actual fault if two or more parameters nominal GARRi in (14) and (15) and the uncertain parts λi
have the same signature and to continue sequential monitor- in (18) and (19). These, correspondingly, give residuals ri and
ing with parameter updating. Because parameter estimation adaptive thresholds ±λi , (i = 3, 4). Inputs to the program
is done locally it consumes less memory and aided with fault are measurements, parameter’s nominal values and uncertainty
sensitivity/direction information, it takes less time to converge. values of parameters of subsystem S1 [10], [28]. Only the mea-
This way, partly decentralized architecture eliminates major surement H 3 of subsystem S2 is required for supervision of
limitations of centralized approach. subsystem S1 . Likewise, subsystem S2 is also monitored inde-
pendently by evaluating the nominal GARRi in (16) and (17)
A. FDI Implementation and the uncertain parts λi in (20) and (21). In this case, only the
In this section, model-based FDI scheme using the most measurement H 2 of subsystem S1 and the nominal parameter
recent existing method [21] which considers single fault Cdv2 are required for supervision of subsystem S2 . To account
hypothesis based on centralized approach and the new pro- for sensor noises, the overall threshold may be evaluated as
posed method which considers sequentially occurring faults ±(λi + ki ), (i = 3, 4, 5, and 6), where ki is the static threshold
hypotheses based on partly decentralized approach, along with which is chosen based on the sensor characteristics or with the
comparison between these two methods are presented. model of additive sensor biases presented in [53] and [54]. In
In particular, we show that the newly proposed method gives the simulations, we have assumed ki = 0.
improved fault isolation capabilities in both single and multiple
faults scenarios when compared with the previously exist- B. Simulation Results
ing methods and also minimizes the computational cost and The four-tank hybrid system is simulated in MATLAB-
requires less memory for computation, hence saving valuable Simulink for duration of 800 s using a fixed step size of 0.02 s
CPU time. and initializing all the state variables to zero. The simulated
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

TABLE V TABLE VII


S IMULATED FAULTS IN THE M ODEL GFSSM (GSS) FOR THE S UBSYSTEM S1

TABLE VI
N OMINAL PARAMETERS OF THE H YBRID T WO -TANK S YSTEM

TABLE VIII
MCSSM FOR THE S UBSYSTEM S1

TABLE IX
GFSSM (GSS) FOR THE S UBSYSTEM S2

faults and nominal system’s parameters used in the model are


given in Tables V and VI, respectively.
The measurements (i.e., H 1 , H 2 , H 3 , H 4 , and QP ) from the
simulated model at a sample rate of 0.02 s are fed to the resid-
uals and thresholds evaluation program for subsystems S1 and
S2 . The time responses of measured outputs (H 1 , H 2 , H 3 , and
H 4 ), input (QP ) and autonomous modes (a1 and a2 ) are shown
in Fig. 11. The transition of modes in subsystem S1 , i.e., a1 TABLE X
and a2 , subjected to different conditions corresponding to two MCSSM FOR THE S UBSYSTEM S2
drainage pipes L1 and L2 , respectively, are clearly noticeable.
For instance, transition of a1 and a2 from 0 to 1 occur at the
time when H 1 (t) and H 2 (t) exceed H L1 and H L2 , respectively.

C. Fault Detection and Isolation


It is observed from the Fig. 11 that within the observa-
tion window, H1 > H2 > H3 > H4 . So the updated GFSSM
and MCSSM for subsystems S1 and S2 during this dura- centralized approach is presented here. Note that we have
tion (0–800 s) are shown in Tables VII–X as static tables. dropped the first two columns of the coherence vector related
In practice, these are variable and updated at every instant. to actuator and controller faults (residuals r1 and r2 ) which
Residuals and thresholds evaluated using the previously are irrelevant in this paper.
existing method (single fault hypothesis based on centralized The coherence vector (C) just after 200 s is obtained
architecture) [21] are given in Fig. 12. Also, the residuals from the evaluated residuals and adaptive thresholds shown
and thresholds evaluated using the newly proposed method in Fig. 12 as C = [1 1 0 0] (if not considering residual sign)
(sequentially occurring faults hypothesis based on partly or C = [−1 +1 0 0] (if considering the residual sign). Note
decentralized architecture) with estimated fault parameters that after 200 s, there is a blockage in V1 (see Table V).
are given in Figs. 13 and 14, for subsystems S1 and S2 , Both these coherence vectors match with the corresponding
respectively. In these figures, solid lines indicate residuals and fault signature of discrete mode fault aV1 given in MCSM
dashed/dotted lines indicate thresholds. (Table II) and MCSSM (Table IV), and Cdv1 given in GFSM
A discussion on fault isolation capabilities using existing (Table I) and GFSSM (Table III), but the match is condition-
FDI methods considering single fault hypothesis based on ally unique. Since last known mode for aV1 = 1; so aV1 = 0
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

OM PRAKASH et al.: MODEL-BASED DIAGNOSIS OF MULTIPLE FAULTS IN HYBRID DYNAMICAL SYSTEMS 13

(a) (b) (a) (b)

(c) (d) (c) (d)

Fig. 11. Time responses of (a) station S1 , (b) station S2 , and (c) pump flow Fig. 12. Residuals (a) r3 , (b) r4 , (c) r5 , and (d) r6 using previously existing
measurements and (d) modes. centralized method [21].

suspected; so, both discrete mode fault aV1 and parametric


fault Cdv1 are suspected. Whether the discrete mode fault in To know the real fault candidates under this scenario,
aV1 occurs or not is checked by evaluating the all ARRs at one needs to simultaneously estimate all system parame-
current mode information and it is found that aV1 is consis- ters by using an appropriate nonlinear parameter estimation
tent. This indicates that the inconsistency is due to parametric technique [20], [45], [47], [48]. However, it is a very difficult
fault of valve V1 . However, from Tables I and IV, valve V1 and time consuming task for a large system by using central-
fault is isolatable only when mode a1 = 0. However, in imme- ized approach; especially considering that quick fault detection
diate aftermath of the fault initiation [see Fig. 11(d)], a1 = 1 and correct fault isolation are the main objectives of the diag-
(āV1 = 0) and thus the fault in valve V1 cannot be isolated. nosis task. However, under the afore-stated assumptions for
Note that in that period, faults in CdV1 (or CdV1 ↓) and CdL1 sequentially occurring faults based on partly decentralized
(or CdL1 ↓) have same signature. While it seems that extending approach, it is required to estimate values of only a few
GFSM to GFSSM apparently does not give any added advan- hypothesized fault candidates [45], [46] or PCs [47]. Least
tage, we will show later that the knowledge of the direction squares response matching or minimization of residual values
of parameter variation as given by GFSSM is very useful in (see [28], [44]–[46] for details) can be used for such param-
expediting the parameter estimation process and indeed the eter estimation. Also, independent subsystems for the local
proposed extensions greatly help in diagnosis. parameter estimation task can be identified by decomposing
When the next blockage fault in V3 is introduced at 300 s, the system model into smaller models [47]. We have used least
the coherence vector just after 300 s is obtained from Fig. 12 as squares response matching approach [43] for local parame-
C = [1 1 1 1] (if not considering residual sign) or C = [−1 +1 ter estimation [47] at the respective smart station with the
−1 +1] (if considering the residual sign). Both these coher- aim of minimum sharing of measurement data among sub-
ence vectors do not match with any of the fault signatures in systems. Such procedure gives better performance when few
Tables I–IV. parameters as matched from GFSSM are to be estimated and
When the next leakage fault in T1 , (CdLeak1 ↑) is introduced bounds (constraints) on the parameter values are known. These
at 400 s, the coherence vector just after 400 s is obtained from constraints on parameter values are implemented as penalty
Fig. 12 as C = [0 1 1 1] (if not considering residual sign) or functions in the objective function used for optimization or
C = [0 +1 −1 +1] (if considering the residual sign). Again, error minimization. As an example, when CdV1 value is to be
both these coherence vectors do not match with any of the estimated, its value can be constrained to lie between a lower
fault signature in Tables I–IV. limit (i.e., 0) and an upper limit. In the considered fault sce-
At last, when the next discrete fault in valve V4 (aV4 ↑) nario (Table V), it is found from GFSSM that the CdV1 has
is introduced at 500 s; the coherence vector just after 500 s possibly decreased and hence, CdV1 value has to lie between
remains same as before and the new fault is not detected by 0 and the earlier known (here, nominal) value. Such con-
using both GFSM and GFSSM. straints reduce the parameter search zone. Also, initial guess
So, detection of faults in both approaches depends upon values for parameters are required for parameter estimation.
the type and sequence of fault occurrence. Sometimes fault is In this case, it may be assigned as αCdV1 , where 0 < α <1
detected or sometimes it is not. Such misdiagnosis is naturally and the residual change rate can be used to approximate the
expected because of the unstructured residuals (i.e., the FSM value of α.
is not diagonal) and hence, cannot be used for diagnosis of A smaller search zone and good initial guess value increase
multiple faults [28]. Thus, we need to use parameter estima- the convergence rate of the optimization algorithm. The param-
tion because the faults cannot be isolated and even detected eter estimation process uses transient process data after fault
sometimes using purely structural analysis approach. occurrence. Thus, there is a delay between the detection and
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

(a) (b)
(a) (b)

(c) (d)
(c) (d)

Fig. 14. (a) and (b) Residuals evaluation by local diagnoser S2 using new
proposed method with post-fault isolation parameter and threshold updating.
(c) Estimated parameters. (d) Locally identified mode.

at smart station S1 obtained through updated nominal part of


(e) (f)
local-GARRs and local updated uncertainties parts obtained
Fig. 13. (a) and (b) Residuals evaluation by local diagnoser S1 using new from the updated local-DHBG model in LFT form (i.e., CdV1
proposed method with post-fault isolation parameter and threshold updating. is replaced by Cf dV1 ) force the residuals to lie within the
(c) Locally identified mode. (d)–(f) Estimated parameters. updated adaptive thresholds (see Fig. 13(a) and (b) just after
time 230 s).
When the next blockage fault in V3 (part of subsystem S2 ),
isolation of a fault. The delay duration includes the time win- is introduced at 300 s, the coherence vector just after 300 s is
dow length of used transient data and the computation time obtained from Fig. 14(a) and (b) as C = [−1 +1]. This coher-
taken for parameter estimation. It is assumed here that no other ence vector matches with fault signatures of discrete mode
fault event occurs during this delay period. Further details fault aV3 ↓ in MCSSM (Table X) and CdV3 ↓ (i.e., blockage
of the estimation process are not given here due to space fault in V3 ) in GFSSM (Table IX). Again mode aV3 is found
constraint. consistent by running all ARR simultaneously [see Fig. 14(d)];
A discussion on improved fault isolatability by using the as a result CdV3 ↓ is detected as a faulty parameter at smart
new proposed FDI method, while considering sequentially station S2 . Thus, again using local parameter estimation with
occurring faults hypotheses based on partly decentralized only possible fault candidate taken as CdV3 , the fault magni-
approach and GFSSM, is presented here. The coherence vec- tude of parameter CdV3 is estimated as Cf dV3 ∼ 0.5CdV3 [see
tor just after 200 s, i.e., after blockage fault initiation in V1 Fig. 14(c) just after 330 s, i.e., 30 s after fault occurrence].
(part of subsystem S1 ), is obtained from the evaluated residu- When this estimated parameter value Cf dV3 is considered as
als checked with adaptive thresholds at local smart station S1 new nominal parameter value in local-DHBG model in LFT
shown in Fig. 13(a) and (b) as C = [−1 +1]. This matches form at smart station S2 to evaluate the residuals and adaptive
with fault signatures of discrete mode fault aV1 ↓ in MCSSM thresholds, the new residuals continue to remain within the
(Table VIII) and CdV1 ↓ (i.e., blockage fault in V1 ) and CdL1 ↓ thresholds [see Fig. 14(a) and (b) just after time 330 s].
(blockage in drainage pipe) in GFSSM (Table VII) when When the next leakage fault in T1 (part of subsystem S1 ),
a1 = 1. Whether the discrete mode fault in aV1 ↓ occurs or is introduced at 400 s, the coherence vector just after 400 s
not is checked by evaluating the all ARRs at current mode is obtained from Fig. 13(a) and (b) as C = [+1 0]. This
information and it is found that mode aV1 is consistent [see coherence vector matches with fault signatures of CdLeak1 ↑
Fig. 13(c)]. This indicates that the inconsistency is due to para- and its magnitude is estimated locally at smart station S1 as
f
metric faults, i.e., CdV1 ↓ or CdL1 ↓. Then the local parameter CdLeak1 ∼ 0.28CdLeak1 . Again, local-DHBG model in LFT
estimation of the suspected parameters CdV1 and CdL1 lead to form at smart station S1 is updated with a new nominal value
isolation of the actual fault and its magnitude, i.e., reported Cf dLeak1 ∼ 0.28CdLeak1 for subsequent fault diagnosis.
by smart station S1 (see Fig. 13(d)–(f) at time just after 230 s At last, when the next discrete fault in valve V4 (aV4 ↑, part
where 30-s time is spent on parameter estimation). In this case, of subsystem S2 ) is introduced at 500 s; the coherence vector
CdV1 is found as actual fault parameter with estimated fault just after 500 s is C = [0 +1], which matches with the signa-
magnitude of Cf dV1 ∼ 0.5CdV1 and this estimated parameter ture of aV4 ↑, CdV4 ↑, and CdLeak4 ↑. Finally aV4 ↑ (stuck on of
value is then onward considered as the new nominal parame- valve V4 ) is identified as a discrete mode fault after simulta-
ter value. Now the residuals and adaptive thresholds evaluation neously running all ARRs with last known mode information
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

OM PRAKASH et al.: MODEL-BASED DIAGNOSIS OF MULTIPLE FAULTS IN HYBRID DYNAMICAL SYSTEMS 15

[see Fig. 14(d)]. This mode information is again used to update level in T4 . Under this scenario, if there is a mode fault due
the local-DHBG model in LFT form at smart station S2 for to which V3 is stuck closed then the controlled state (water
subsequent fault diagnosis. level in T4 ) cannot be reached by the control input and the
This way, subsequent discrete and parametric faults can be integral gain tries to make the system unstable, but cannot do
correctly isolated if process operation may be continued after so because of the saturation of the pump input (10) and the
detection and isolation of minor faults. In this system, tran- drainage line present in T2 . Therefore, the system under con-
sient data for 27 s is used for the parameter estimation and sideration is so chosen/ configured that it remains stable and
discrete fault identification. The parameter estimation takes controllable under all parametric and mode faults.
less than 3 s in a standard personal computer. Thus, the esti-
mation window length is 30 s. Discrete mode fault information V. E QUIVALENT E LECTRICAL S YSTEM
or parameters and the thresholds are updated after 30 s. This
A linear hydraulic system can be represented by an anal-
window length mostly depends on the dynamic system’s time
ogous linear electrical system with similar dynamics or gov-
constant.
erning differential equations. For the experimental validation,
This method gives better FDI capabilities and takes less
equivalent electrical system is modeled for a benchmark two-
CPU time as compared with existing centralized diagnosis
tank hybrid system adapted from [10] to show the robustness
approach that does not take the advantage of direction of
of proposed method. This two-tank hybrid system is similar
fault information (GFSSM/MCSSM) and uses global model
to the subsystem S1 of the considered four tank hybrid sys-
for parameter estimation. The proposed approach is also capa-
tem with minor changes. In this case, the flow through valve
ble of isolating actual fault even when fault effects may be
V2 goes to surrounding in place of tank T3 by disconnect-
concealed or compensated by the effects of other faults as
ing the valve V2 from tank T3 . In this revised model, we
in Fig. 12(a) where residual r3 hides fault in valve 1 in pres-
will show how the targeted parameter estimation and updating
ence of leakage from tank 1 after 400 s. The proposed method
by using fault direction information can avoid fault mask-
reduces the list of fault candidates or PCs, uses local param-
ing while saving valuable CPU time in comparison to full
eter estimation with minimum information exchange among
parameter estimation with no prior information of parame-
subsystems, and helps in setting proper limits/bounds/penalty
ter deviation direction. For building the equivalent electrical
for parameter values and guess initial parameter values dur-
system of the hybrid two-tank system, it is assumed that the
ing parameter estimation process. Also, it is scalable to large
hydraulic flow is incompressible and driven by pressure dif-
systems through decoupling of the GFSSM and MCSSM of
ference with low Reynolds number so that the inertia effect
global model to manageable local diagnosis models.
of the fluid is neglected. The equivalent electrical system is
further scaled down for easy and low cost experimental imple-
D. Stability of the System mentation. Note that while the equivalent electrical system is
linear, the hybrid nature (mode dependent changes) will be
Along with FDI, FTC may be implemented in order to
retained in it.
activate a proper controller among several controllers to
accommodate any discrete mode fault or large parametric fault
(if possible) so as to guarantee the global stability and per- A. Circuit Layout
formance objectives of the overall system. The open loop Equivalent electrical circuit of the considered two-tank sys-
four-tank system is BIBO stable in any mode because it only tem is presented in Fig. 15. In the electrical domain, power
contains storage elements (C elements) and positive dissipa- variables voltage and current, respectively, are equivalent to
tion (R elements) with no indefinite energy import from active the pressure and flow rate in hydraulic domain. In the circuit,
sources. The only cause of unbounded input can be from feed- two electrical capacitors C1 and C2 are used in place of two
back instability, especially the PI controller used in the system. tanks T1 and T2 , respectively. Resistors R1 , R2 and Rd1 , Rd2
Large gains in the PI controller can make certain systems are used for valves V1 and V2 and drainage pipes L1 and L2 ,
unstable; but not so in this case. For the considered system, respectively. Diodes D1 and D2 are used as switches to allow
linearization at any operating point yields a transfer function the current in equivalent drainage pipe resistors Rd1 and Rd2
with all poles and zeroes on the real axis, and all these poles at certain set threshold voltages V set1 and V set2 , respectively.
and zeroes are on the left half of s-plane. A PI controller Current source I d1 is the drainage current through resistor Rd1
adds a pole at origin and a zero on negative real axis. As which charges the capacitor C2 as per the corresponding set
a consequence, the number of asymptotes of root loci does condition.
not change. The system remains phase minimal at all times. In place of modulated pump flow QP , a modulated current
Also for partial parametric faults and mode faults, the con- source I in is used. The modulated current source I in is achieved
trolled state (the water level in tank 1) is reachable from the by using a known resistor R and controlling the input voltage
control input and the system is structurally controllable as per V in , which is the output of the PI-controller. The PI-controller
the conditions given in [59]. However, it should be noted that is constructed using different OP-AMPs whose output is the
a slight change in the system configuration can lead to an modulated voltage V in . The aim of PI-controller is to maintain
uncontrollable and also unstable performance under a mode a desired set voltage (V set ) across the capacitor C1 (equivalent
fault. For example, consider that the PI controlled pump sup- to level set point in tank T1 ). Two switches Sw1 and Sw2 are
plies flow to T1 whereas the controlled variable is the water used to introduce the faults in resistors R1 and R2 , respectively.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

16 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

TABLE XI
N OMINAL PARAMETERS OF THE S CALED
E QUIVALENT E LECTRICAL C IRCUIT

Fig. 15. Schematic of equivalent electrical circuit of two-tank.

Due to parallel connection, the resistance is low when the


switch is on (nominal state) and high when it is off (simulated
faulty state). Sw3 is also used to make the resistor R1 on and off B. Model Scaling
in the circuit; the state of Sw3 is on throughout the experiment
For experiment, the voltage and capacitance values obtained
time. The level sensors (H 1 , H 2 ) and flow sensor (QP ) in the
from direct scaling were found to be out of range. For example,
hydraulic circuit are replaced by voltage sensors (V s1 , V s2 )
as per (24), we need to operate around 5-kV electrical voltage
and current sensor (I in ), respectively, in the electrical domain.
range. Thus, a dual time and amplitude scaling was performed.
The PI-controller law used in the modulated current source is
For linear systems, the response x(t) of the original system may
given as
be scaled as

Vin (t) = KP (Vset − Vs1 (t)) + KI (Vset1 − Vs1 (t))dt (22) xs (ts ) = αx(βt) (26)
Vin (t) − Vs1 (t)
Iin (t) = . (23) where α and β are, respectively, the constants for amplitude
R and time scaling. We have used α = 1/981 which means the
The values of the pressures P1 and P2 and pump input flow input current is reduced or equivalently, the voltage set point
QP of the adapted hydraulic two-tank system at steady state for PI controller is set at 4905/981 = 5 V.
were measured by doing a simulation in MATLAB-Simulink The hydraulic system has a slow response. We reduced the
using the same parameters values as shown in Table VI. These response time by β = 10 times. For that, the time constant
steady-state values are (τ = RC) was adjusted by retaining the value of R and reduc-
ing the value of C by ten times. The gains of the PI controller
P1 = ρ · g · H1 = 1000 × 9.81 × 0.5 = 4905 N/m2 are adjusted to match with time constant of the modified sys-
P2 = ρ · g · H2 = 1000 × 9.81 × 0.25 = 2452.5 N/m2 tem. The final scaled parameter values given in Table XI are
used for developing the experimental setup.
QP = 0.8 kg/s. (24)

The nominal parameters of the equivalent electrical circuit VI. E XPERIMENTAL S TUDY
are obtained by hydraulic-electrical analogy. For instance, the
The experimental test-bed is presented in Fig. 16(a). The
equivalent linear electrical resistances corresponding to the
circuit in the breadboard is the equivalent electrical hybrid
two valves V1 and V2 are obtained as
system of the simulated two-tank system in reduced scale
P1 − P2 4905 − 2452.5 whose dynamics is approximately similar to the two tank sys-
R1 = = = 3065.625 tem dynamics. The schematic diagram of equivalent electrical
QP 0.8
P2 2452.5 circuit of two-tank system is already presented in Fig. 15.
R2 = = = 3065.625 . (25) The various parts of the circuit including the PI controller are
QP 0.8
marked in Fig. 16(b).
Time constant matching was used to obtain the values of The experimental data are collected at a fixed sampling
the capacitances. The rounded off values of parameters of the rate of 0.02 s from voltage sensors V s1 and V s2 and cur-
equivalent electrical system are obtained as follows: rent sensor I in using a data-acquisition card (NI-USB6211).
LabVIEW-MATLAB interface was used for computation of
R = R1 = R2 = 3066 , Rd1 = Rd2 = 1000 residuals, thresholds and parameter estimation. The simulation
C1 = C2 = 1500 μF. results from the equivalent electrical circuit and the experimen-
tal results from the setup are compared in Fig. 17. A good
The resistances, capacitances, op-amps, and other electri- agreement is obtained between responses of simulated and
cal/electronic components procured from market come with experimental systems. This validated setup is used further for
some uncertainties. Thus, the rounding-off errors may be testing the developed diagnosis algorithms for sequentially
treated as parametric uncertainties. occurring faults.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

OM PRAKASH et al.: MODEL-BASED DIAGNOSIS OF MULTIPLE FAULTS IN HYBRID DYNAMICAL SYSTEMS 17

TABLE XII
GFSSM (GSS) FOR THE H YBRID E LECTRICAL S YSTEM

TABLE XIII
MCSSM FOR THE H YBRID E LECTRICAL S YSTEM

TABLE XIV
I NTRODUCED FAULTS IN THE E XPERIMENT

Fig. 16. (a) Test-bed of electrical hybrid system. (b) Enlarge view of bread-
board showing different components with dotted box: 1) PI-controller; 2) R;
3) C1 ; 4) C2 ; 5) R1 ; 6) R2 ; 7) Sw1 ; 8) Sw2 ; 9) I d1 ; 10) D1 , Rd1 , and V set1 ,
11) D2 , Rd2 , and V set2 ; and 12) buffers.

as shown in Fig. 18(a) and (b). A low-pass filter (with a time


constant 0.1 s) is used in order to eliminate unwanted noise
from the residuals (r3 and r4 ).
From Fig. 18(a) and (b), the coherence vector just after 60 s
is C = [0, −1] with two possible fault candidates (R2 ↑ and
Rd2 ↑) for mode a2 = 1 in GFSSM (Table XII). Upon param-
eter estimation [see Fig. 18(c) and (d)], the fault was found
to be R2 ↑. Note that here the process has a faster response
(a) (b)
for which only 2.7 s transient data after fault detection is
Fig. 17. (a) Voltages and (b) Current measurements from simulation and needed for parameter estimation. The computational time for
experiment. parameter estimation is same as that for the hydraulic sys-
tem (about 3 s). Thus, the parameter and threshold updating
The measured data are fed into the DHBG model of the elec- is delayed by 5.7 s after every fault occurrence. Thereafter,
trical hybrid system in LFT form for residuals and adaptive the residuals computed with updated parameter value remain
thresholds evaluation. The GFSSM and MCSSM of the equiv- within the bounds of the updated thresholds. When the next
alent electrical hybrid system given in Tables XII and XIII fault is introduced at 120 s, the coherence vector is C = [−1,
are same as hydraulic two-tank system or subsystem S1 ; only +1]. This matches with fault signatures of discrete mode fault
the corresponding nomenclatures of parameters are changed. aR1 ↓ in MCSSM (Table XIII), and R1 ↑ and Rd1 ↑ in GFSSM
For instance, the coefficient of discharge parameter Cdi ↓ is (Table XII) when a1 = 1. Whether the discrete mode fault in
equivalent to 1/Ri ↓ where Ri is the electrical linear resistance, aR1 occurs or not is checked by evaluating the all ARRs using
i.e., the signature of Cdi ↓ is replaced by Ri ↑. Similarly, the previous known mode information and it is found that mode
signature of Cdi ↑ is replaced by Ri ↓ and so on. aR1 is consistent. This indicates that the inconsistency is due to
The introduced faults in the parameters R1 and R2 and their parametric faults, i.e., R1 ↑ or Rd1 ↑. The real fault is isolated
occurrences in the experiment are given in Table XIV. as R1 ↑ from parameter estimation [see Fig. 18(e) and (f)].
In a similar manner as in simulation (Section IV-C), sequen- Note that there is a mode change in a1 after a short time from
tially occurring faults are detected and isolated by the new introduction of the second fault [similar to the mode change
proposed method. The residuals and adaptive thresholds were shown in Fig. 11(d) just after 200 s]. Thus, the coherence vec-
evaluated from the real measurement data from the experiment tor C = [−1, +1] is found for a1 = 1 and a1 = 0. As a result,
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

18 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

of fault parameters. It was found that richer qualitative behav-


ior information extracted from sensitivity signatures allows
quicker and reliable identification of faults. After estimation of
the fault parameter, the diagnosis model and adaptive residual
thresholds are updated so that the faulty system or subsystem
can be treated as a normally operating system or subsystem
with some known degradation (as a new nominal model). For
(a) (b) a system or subsystem with few sensors and unstructured (non-
diagonal) residuals, this approach removes the possibility of
any further fault effect being masked due to previously exist-
ing faults and thus is able to diagnose subsequent faults. The
developed method is first applied to a simulated four tank
hybrid system and then it was experimentally validated on
a scaled equivalent electrical circuit model of a two tank hybrid
(c) (d) system.
The method proposed in this paper reduces the list of PCs
and aided by fault direction information, it tries to speed up
the parameter estimation process. Also, by updating the new
parameter value in the nominal model, subsequent faults can
be isolated. The response time (time constant) of the system
and the time taken for parameter estimation are two criti-
cal parameters that govern the applicability of this diagnosis
(e) (f) scheme. For processes with slow response time, the time spent
for parameter estimation is insignificant. However, for sys-
Fig. 18. Residuals (a) r3 and (b) r4 using new proposed method, and tems with fast dynamics, the parameter estimation needs to
estimated parameters (c) R2 , (d) Rd2 , (e) R1 and (f) Rd1 .
be faster which can be achieved through distributed diagno-
sis and increased number of sensors to localize the initially
R1 ↑ is directly isolated as the fault. However, estimation of hypothesized fault more precisely.
the fault magnitude still requires parameter estimation and the This paper does not directly deal with fault accommoda-
data used for it contains a mode change. When the estimated tion (recovery from a fault) issues. If a fault is accommodated
parameter value is finally updated as the nominal parameter through FTC by changing the control laws and maintaining the
value at 125.7 s (after 5.7 s delay), the revised residuals and system stability then the proposed diagnosis scheme does not
thresholds indicate normal process operation. Further faults require any modification. One example is increasing the water
occurring in the circuit, including further changes in param- level set point in tank T1 in order to maintain the output flow
eter values of the already faulty components, can be isolated in the case of a blockage fault in valve V1 . On the other hand,
in similar way. if the system is reconfigured (e.g., a parallel line of flow is
activated) then a new system architecture results and the FDI
system has to be reinitialized with the new set of estimated
VII. C ONCLUSION parameter values, GARRS, GFSSM, MCSSM, etc.
The problem of detection and isolation of sequentially Fault estimation and reconstruction in hybrid dynamical
occurring faults (considering both discrete and parametric systems is complicated because the mode information has
faults) based on partly decentralized diagnosis architecture to be considered in the algorithm. If the residual incon-
for hybrid systems are presented in this paper. The diagnosis sistency is confirmed to be not due to a mode fault then
scheme uses system measurements in the transient regime for it is assumed to be due to a parametric fault and param-
discrete mode fault identification and parameter estimation. eter estimation is required to estimate the parametric fault
It is assumed that a single fault occurs at a time and there magnitude with known mode information. For parameter esti-
is a specific minimum time gap between the occurrences of mation, sampled data of a fixed and proper time window is
two sequential faults. This time gap is determined from the needed.
system response time and the computation time needed for 1) If the parameter fault is related to a hybrid compo-
parameter estimation process and discrete mode fault iden- nent (e.g., blockage fault during on state of an on–off
tification. GFSSM and MCSSM are used to make a local valve) then the collected measurement data of a fixed
diagnoser for a global large system with minimum exchange of time window should belong to the on state of the valve.
measurements among the local diagnosers. The possible fault Otherwise, parameter estimation provides wrong results.
directions are identified from the residual responses by using If the state of the faulty valve changes from on state to
the Global fault sensitivity signatures of the hybrid dynamical off state during sample data collection for some reason
system. The information of possible fault candidates and the then new fresh sample data of a fixed window size can
directions of corresponding parameter deviations along with only be collected after the same valve reverts back to
operating mode information are used for targeted estimation on state.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

OM PRAKASH et al.: MODEL-BASED DIAGNOSIS OF MULTIPLE FAULTS IN HYBRID DYNAMICAL SYSTEMS 19

2) If the parameter is not related to a hybrid com- R EFERENCES


ponent then only the correct mode information of [1] V. Venkatasubramanian, R. Rengaswamy, K. Yin, and S. N. Kavuri,
system is needed for parameter estimation of the “A review of process fault detection and diagnosis: Part I: Quantitative
model-based methods,” Comput. Chem. Eng., vol. 27, no. 3,
faulty component. There is no need to collect data for pp. 293–311, 2003.
a frozen mode; it can be collected over any time window [2] V. Venkatasubramanian, R. Rengaswamy, K. Yin, and S. N. Kavuri,
even when modes are arbitrarily changing. Of course, the “A review of process fault detection and diagnosis: Part II: Qualitative
models and search strategies,” Comput. Chem. Eng., vol. 27, no. 3,
mode change information has to be also logged with the pp. 313–326, 2003.
time series data. [3] V. Venkatasubramanian, R. Rengaswamy, S. N. Kavuri, and K. Yin,
“A review of process fault detection and diagnosis: Part III: Process his-
3) The estimation of faults in storage element parameters tory based methods,” Comput. Chem. Eng., vol. 27, no. 3, pp. 327–346,
always requires transient data, i.e., data before the sys- 2003.
[4] Z. Gao, C. Cecati, and S. X. Ding, “A survey of fault diagnosis and
tem reaches a steady state. This is because in derivative fault-tolerant techniques—Part I: Fault diagnosis with model-based and
causality, the constitutive relation of a storage element signal-based approaches,” IEEE Trans. Ind. Electron., vol. 62, no. 6,
pp. 3757–3767, Jun. 2015.
involves time derivative of the signal and in steady state, [5] Z. Gao, C. Cecati, and S. X. Ding, “A Survey of fault diagnosis and
the time derivative becomes trivially zero. fault-tolerant techniques—Part II: Fault diagnosis with knowledge-based
4) Knowing the possible fault candidates, implementing the and hybrid/active approaches,” IEEE Trans. Ind. Electron., vol. 62, no. 6,
pp. 3768–3774, Jun. 2015.
proper time window for data collection is a challenging [6] I. Roychoudhury, G. Biswas, and X. Koutsoukos, “Designing distributed
task, especially when the system size increases. diagnosers for complex continuous systems,” IEEE Trans. Autom. Sci.
Eng., vol. 6, no. 2, pp. 277–290, Apr. 2009.
Requirement of slow dynamics (response time) of the sys- [7] E. Chanthery, L. Travé-Massuyés, and S. Indra, “Fault isolation on
tem, such as in a process engineering plant, is a limitation request based on decentralized residual generation,” IEEE Trans. Syst.,
Man, Cybern., Syst., vol. 46, no. 5, pp. 598–610, May 2016.
of the presented approach. The developed approach also fails [8] M. J. Daigle, X. D. Koutsoukos, and G. Biswas, “An event-based
when two or more faults occur at the same time or with very approach to integrated parametric and discrete fault diagnosis in hybrid
little time gap. For the hydraulic system considered in this systems,” Trans. Inst. Meas. Control, vol. 32, no. 5, pp. 487–510, 2009.
[9] O. Prakash and A. K. Samantaray, “Model-based diagnosis and prog-
paper, the minimum fault event separation is 30 s whereas for nosis of hybrid dynamical systems with dynamically updated parame-
the experimented electrical system, it is 5.7 s. For FDI imple- ters,” in Bond Graphs for Modelling, Control and Fault Diagnosis of
Engineering Systems. Cham, Switzerland: Springer, 2017, pp. 195–232.
mentation in a new system, these separation times must be [10] S. K. Ghoshal, S. Samanta, and A. K. Samantaray, “Robust fault detec-
known a priori either through simulations or experiments. tion and isolation of hybrid systems with uncertain parameters,” Proc.
Inst. Mech. Eng. I J. Syst. Control Eng., vol. 226, no. 8, pp. 1013–1028,
Another requirement is that the system must be 2012.
monotone [60] with respect to hypothesized fault parameters, [11] C. B. Low, D. Wang, S. Arogeti, and J. B. Zhang, “Causality assign-
i.e., for a unique pair of input-output sequence there exist ment and model approximation for hybrid bond graph: Fault diagnosis
perspectives,” IEEE Trans. Autom. Sci. Eng., vol. 7, no. 3, pp. 570–580,
a unique set of parameter values. This condition guarantees Jul. 2010.
that from the set of hypothesized fault candidates, only one [12] C. B. Low, D. Wang, S. Arogeti, and M. Luo, “Quantitative hybrid bond
graph-based fault detection and isolation,” IEEE Trans. Autom. Sci. Eng.,
specific parameter combination can give the same response vol. 7, no. 3, pp. 558–569, Jul. 2010.
in the observed time window after the occurrence of the [13] R. Levy, S. A. Arogeti, and D. Wang, “An integrated approach to mode
tracking and diagnosis of hybrid systems,” IEEE Trans. Ind. Electron.,
suspected fault (hypothesis validation). Unless this condition vol. 61, no. 4, pp. 2024–2040, Apr. 2014.
is satisfied, there can be false detections of fault. This [14] D. Wang, M. Yu, C. B. Low, and S. Arogeti, Model-Based Health
Monitoring of Hybrid Systems. New York, NY, USA: Springer, 2013.
condition may be satisfied under certain modes and not [15] W. Borutzky, “Bond-graph-based fault detection and isolation for hybrid
satisfied in certain other modes. Fortunately, this condition system models,” Proc. Inst. Mech. Eng. I J. Syst. Control Eng., vol. 226,
holds for most class of dynamical physical systems [61]. In no. 6, pp. 742–760, 2012.
[16] W. Borutzky, “Bond graph modelling and simulation of fault scenarios
the examples in this paper, the monotonicity condition is in switched power electronic systems,” Proc. Inst. Mech. Eng. I J. Syst.
satisfied in all modes except one. The only situation when Control Eng., vol. 226, no. 10, pp. 1381–1393, 2012.
[17] S. Narasimhan and G. Biswas, “Model-based diagnosis of hybrid sys-
the hydraulic system (and the equivalent electrical system) tems,” IEEE Trans. Syst., Man, Cybern. A, Syst., Humans, vol. 37, no. 3,
does not satisfy monotonicity is the case where valve V4 is pp. 348–361, May 2007.
[18] S. A. Arogeti, D. Wang, C. B. Low, and M. Luo, “Energy-based mode
in on state and the possible fault candidates are leakage from tracking of hybrid systems for FDI,” IEEE Trans. Syst., Man, Cybern.,
tank T4 and leakage from valve V4 . Physically, leakage from Syst., vol. 43, no. 1, pp. 14–28, Jan. 2013.
tank T4 and valve V4 are indistinguishable from each other; [19] S. A. Arogeti, D. Wang, and C. B. Low, “Mode identification of hybrid
systems in the presence of fault,” IEEE Trans. Ind. Electron., vol. 57,
the effective leakage flow depends on the combined flow no. 4, pp. 1452–1467, Apr. 2010.
rate and there are infinite number of parametric (discharge [20] W. Borutzky, “Isolation of multiple parametric faults from a hybrid
model,” in Bond Graph Model-Based Fault Diagnosis of Hybrid Systems.
coefficient) combinations of the two that give the same result. Cham, Switzerland: Springer, 2015, pp. 123–148.
The monotonicity condition can be restored if a flow sensor [21] R. Levy, S. Arogeti, D. Wang, and O. Fivel, “Improved diagnosis of
hybrid systems using instantaneous sensitivity matrices,” Mech. Mach.
is added at valve V4 , which would also uniquely isolate the Theory, vol. 91, pp. 240–257, Sep. 2015.
individual faults (from modified GFSSM) without resorting [22] A. K. Samantaray, S. K. Ghoshal, and S. Chakraborty, “Bond graph
to parameter estimation. Monotonicity of a dynamical system model based design of supervision algorithm for distributed fault tolerant
control systems,” Int. J. Autom. Control, vol. 1, no. 1, pp. 28–47, 2007.
can be ascertained from inspection of its system equations to [23] A. Bregon et al., “An event-based distributed diagnosis framework
identify identical dynamics or through numerical simulations. using structural model decomposition,” Artif. Intell., vol. 210, pp. 1–35,
May 2014.
In conclusion, monotonicity of the dynamical system with [24] A.-L. Gehin and B. O. Bouamama, “Design of distributed fault detec-
respect to the set of hypothesized fault candidates is a tion and isolation systems,” IFAC PapersOnLine, vol. 48, no. 21,
pp. 907–912, 2015.
precondition for application of the diagnosis scheme proposed [25] D. C. Karnopp, D. L. Margolis, and R. C. Rosenberg, System Dynamics:
in this paper. A Unified Approach, 2nd ed. New York, NY, USA: Wiley, 1990.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

20 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

[26] A. Mukherjee, R. Karmakar, and A. K. Samantaray, Bond Graph [53] Y. Touati, R. Merzouki, and B. O. Bouamama, “Robust diagnosis to mea-
in Modelling, Simulation and Fault Identification. New Delhi, India: surement uncertainties using bond graph approach: Application to intelli-
I. K. Int., 2006. gent autonomous vehicle,” Mechatronics, vol. 22, no. 8, pp. 1148–1160,
[27] W. Borutzky, Bond Graph Methodology: Development and Analysis 2012.
of Multidisciplinary Dynamic System Models. Londan, U.K.: Springer, [54] W. Borutzky, “Parameter uncertainties,” in Bond Graph Model-Based
2010. Fault Diagnosis of Hybrid Systems. Cham, Switzerland: Springer, 2015,
[28] A. K. Samantaray and B. O. Bouamama, Model-Based Process pp. 101–121.
Supervision: A Bond Graph Approach. London, U.K.: Springer, 2008. [55] Z. Gao, X. Dai, T. Breikin, and H. Wang, “Novel parameter identification
[29] B. O. Bouamama, K. Medjaher, A. K. Samantaray, and M. Staroswiecki, by using a high-gain observer with application to a gas turbine engine,”
“Supervision of an industrial steam generator. Part I: Bond graph IEEE Trans. Ind. Informat., vol. 4, no. 4, pp. 271–279, Nov. 2008.
modelling,” Control Eng. Pract., vol. 14, no. 1, pp. 71–83, 2006. [56] Z. Gao, T. Breikin, and H. Wang, “Reliable observer-based control
[30] A. K. Samantaray, K. Medjaher, B. O. Bouamama, M. Staroswiecki, and against sensor failures for systems with time delays in both state and
G. Dauphin-Tanguy, “Diagnostic bond graphs for online fault detec- input,” IEEE Trans. Syst., Man, Cybern. A, Syst., Humans, vol. 38, no. 5,
tion and isolation,” Simulat. Model. Pract. Theory, vol. 14, no. 3, pp. 1018–1029, Sep. 2008.
pp. 237–262, 2006. [57] M. Ji, Z. Zhang, G. Biswas, and N. Sarkar, “Hybrid fault adaptive control
[31] W. Borutzky, “Bond graph model-based fault detection using residual of a wheeled mobile robot,” IEEE/ASME Trans. Mechatronics, vol. 8,
sinks,” Proc. Inst. Mech. Eng. J. Syst. Control Eng., vol. 223, no. 3, no. 2, pp. 226–233, Jun. 2003.
pp. 337–352, 2009. [58] B. Jiang, H. Yang, and V. Cocquempot, “Results and perspectives on
[32] G. Biswas, M.-O. Cordier, J. Lunze, L. Trave-Massuyes, and fault tolerant control for a class of hybrid systems,” Int. J. Control,
M. Staroswiecki, “Diagnosis of complex systems: Bridging the method- vol. 84, no. 2, pp. 396–411, 2011.
ologies of the FDI and DX communities,” IEEE Trans. Syst., Man, [59] A. Rahmani and G. Dauphin-Tanguy, “Structural analysis of switching
Cybern. B, Cybern., vol. 34, no. 5, pp. 2159–2162, Oct. 2004. systems modelled by bond graph,” Math. Comput. Model. Dyn. Syst.,
[33] A. Bregon, G. Biswas, B. Pulido, C. Alonso-Gonzalez, and vol. 12, nos. 2–3, pp. 235–247, 2006.
H. Khorasgani, “A common framework for compilation techniques [60] D. Angeli and E. D. Sontag, “Monotone control systems,” IEEE Trans.
applied to diagnosis of linear dynamic systems,” IEEE Trans. Syst., Man, Autom. Control, vol. 48, no. 10, pp. 1684–1698, Oct. 2003.
Cybern., Syst., vol. 44, no. 7, pp. 863–876, Jul. 2014. [61] H. L. Smith, “Monotone dynamical systems: An introduction to the the-
[34] P. J. Mosterman and G. Biswas, “Behavior generation using model ory of competitive and cooperative systems,” in Mathematical Surveys
switching: A hybrid bond graph modelling technique,” Trans. Soc. and Monographs, vol. 41. Providence, RI, USA: Amer. Math. Soc.,
Comput. Simulat., vol. 27, no. 1, pp. 177–182, 1995. 1995.
[35] I. Roychoudhury, M. J. Daigle, G. Biswas, and X. Koutsoukos, “Efficient
simulation of hybrid systems: A hybrid bond graph approach,” Simulat.
Trans. Soc. Model. Simulat. Int., vol. 87, no. 6, pp. 467–498, 2011. Om Prakash received the B.Tech. degree from
[36] M. J. Daigle et al., “A comprehensive diagnosis methodology for com- Ranchi University, Ranchi, India, in 2009, and
plex hybrid systems: A case study on spacecraft power distribution the master’s degree from the National Institute
systems,” IEEE Trans. Syst., Man, Cybern. A, Syst., Humans, vol. 40, of Technology, Jamshedpur, Jamshedpur, India, in
no. 5, pp. 917–931, Sep. 2010. 2011, both in mechanical engineering.
[37] M. Staroswiecki and G. Comtet-Varga, “Analytical redundancy rela-
tions for fault detection and isolation in algebraic dynamic systems,” He was a Senior Engineer with the Product
Automatica, vol. 37, no. 5, pp. 687–699, 2001. Engineering Department, TRF Ltd., Jamshedpur,
[38] Z. Shi, F. Gu, B. Lennox, and A. D. Ball, “The development of an from 2011 to 2013. Since 2013, he has been
adaptive threshold for model-based fault detection of a nonlinear electro- a Doctoral Researcher of Model-Based Fault
hydraulic system,” Control Eng. Pract., vol. 13, no. 11, pp. 1357–1367, Diagnosis and Prognosis with the Systems,
2005. Dynamics and Control Laboratory, Mechanical
[39] M. A. Djeziri, R. Merzouki, B. O. Bouamama, and G. Dauphin-Tanguy, Engineering Department, Indian Institute of Technology Kharagpur,
“Robust fault diagnosis by using bond graph approach,” IEEE/ASME Kharagpur, India. His current research interests include bond graph
Trans. Mechatronics, vol. 12, no. 6, pp. 599–611, Dec. 2007. modelling, simulation, and fault diagnosis.
[40] M. A. Djeziri, B. O. Bouamama, and R. Merzouki, “Modelling and
robust FDI of steam generator using uncertain bond graph model,”
J. Process Control, vol. 19, no. 1, pp. 149–162, 2009.
Arun K. Samantaray received the B.Tech. degree
[41] W. Borutzky, “Bond graph model-based system mode identification and
mode-dependent fault thresholds for hybrid systems,” Math. Comput. from CET, Bhubaneswar, Bhubaneswar, India, in
Model. Dyn. Syst., vol. 20, no. 6, pp. 584–615, 2014. 1989, and the M.Tech. degree in machine dynam-
[42] P. M. Frank, Introduction to System Sensitivity Theory. New York, NY, ics and the Ph.D. degree from the Indian Institute of
USA: Academic Press, 1978. Technology Kharagpur (IIT Kharagpur), Kharagpur,
[43] P. J. Gawthrop, “Sensitivity bond graphs,” J. Frankl. Inst., vol. 337, India, in 1991 and 1996, respectively.
no. 7, pp. 907–922, 2000. He is a Full Professor and in-charge of
[44] A. K. Samantaray and S. K. Ghoshal, “Bicausal bond graphs for super- the Systems, Dynamics and Control Laboratory,
vision: From fault detection and isolation to fault accommodation,” Mechanical Engineering Department, IIT Kharagpur.
J. Frankl. Inst., vol. 345, no. 1, pp. 1–28, 2008. He was a Project Manager for HighTech consul-
[45] A. K. Samantaray and S. K. Ghoshal, “Sensitivity bond graph approach tants from 1996 to 2000, and a CNRS Researcher
to multiple fault isolation through parameter estimation,” Proc. Inst. at Université Lille 1: Sciences et Technologies, Villeneuve-d’Ascq, France,
Mech. Eng. J. Syst. Control Eng., vol. 221, no. 4, pp. 577–587, 2007. from 2001 to 2004. He has authored three books and many journal articles.
[46] A. K. Samantaray, S. K. Ghoshal, S. Chakraborty, and A. Mukherjee, He is also a consultant to various industries. His current research interests
“Improvements to single-fault isolation using estimated parameters,”
Simulat. Trans. Soc. Model. Simulat. Int., vol. 81, no. 12, pp. 827–845, include systems and control, fault diagnosis, automation, nonlinear mechanics,
2005. robotics, rotor dynamics, and vehicle dynamics.
[47] A. Bregon, G. Biswas, and B. Pulido, “A decomposition method for
nonlinear parameter estimation in TRANSCEND,” IEEE Trans. Syst.,
Man, Cybern. A, Syst., Humans, vol. 42, no. 3, pp. 751–763, May 2012. Ranjan Bhattacharyya received the B.Tech. and
[48] C. B. Low, D. Wang, S. A. Arogeti, and M. Luo, “Fault parameter esti- M.Tech. degrees from the Indian Institute of
mation for hybrid systems using hybrid bond graph,” in Proc. 3rd IEEE Technology Kharagpur (IIT Kharagpur), Kharagpur,
Conf. Syst. Control, St. Petersburg, Russia, Jul. 2009, pp. 1338–1343. India, in 1981 and 1983, respectively, and the Ph.D.
[49] E. Frisk et al., “Diagnosability analysis considering causal interpreta- degree from the University of Kentucky, Lexington,
tions for differential constraints,” IEEE Trans. Syst., Man, Cybern. A, KY, USA, in 1987.
Syst., Humans, vol. 42, no. 5, pp. 1216–1229, Sep. 2012. He is a Full Professor with the Mechanical
[50] M. Yu and J. Xu, “Sequential fault diagnosis for mechatronics system
using diagnostic hybrid bond graph and composite harmony search,” Engineering Department, IIT Kharagpur, where
Adv. Mech. Eng., vol. 7, no. 12, pp. 1–14, 2015. he was the Head of the Mechanical Engineering
[51] R. Goebel, R. G. Sanfelice, and A. R. Teel, “Hybrid dynamical systems,” Department, from 2010 to 2013. He was a Visiting
IEEE Control Syst., vol. 29, no. 2, pp. 28–93, Apr. 2009. Assistant Professor with the University of Pittsburgh
[52] W. Borutzky, “Incremental bond graphs,” in Bond Graph Modelling at Bradford, Bradford, PA, USA, till 1990. His current research inter-
of Engineering Systems. New York, NY, USA: Springer, 2011, ests include nonlinear dynamics, nonlinear elasticity, and control systems
pp. 135–176. engineering.

You might also like